2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, you can access it online at
16 * http://www.gnu.org/licenses/gpl-2.0.html.
18 * Copyright IBM Corporation, 2001
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
23 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
24 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
26 * http://www.rdrop.com/users/paulmck/paper/rclockpdcsproof.pdf
27 * http://lse.sourceforge.net/locking/rclock_OLS.2001.05.01c.sc.pdf (OLS2001)
29 * For detailed explanation of Read-Copy Update mechanism see -
30 * http://lse.sourceforge.net/locking/rcupdate.html
33 #include <linux/types.h>
34 #include <linux/kernel.h>
35 #include <linux/init.h>
36 #include <linux/spinlock.h>
37 #include <linux/smp.h>
38 #include <linux/interrupt.h>
39 #include <linux/sched/signal.h>
40 #include <linux/sched/debug.h>
41 #include <linux/atomic.h>
42 #include <linux/bitops.h>
43 #include <linux/percpu.h>
44 #include <linux/notifier.h>
45 #include <linux/cpu.h>
46 #include <linux/mutex.h>
47 #include <linux/export.h>
48 #include <linux/hardirq.h>
49 #include <linux/delay.h>
50 #include <linux/moduleparam.h>
51 #include <linux/kthread.h>
52 #include <linux/tick.h>
53 #include <linux/rcupdate_wait.h>
54 #include <linux/sched/isolation.h>
56 #define CREATE_TRACE_POINTS
60 #ifdef MODULE_PARAM_PREFIX
61 #undef MODULE_PARAM_PREFIX
63 #define MODULE_PARAM_PREFIX "rcupdate."
65 #ifndef CONFIG_TINY_RCU
66 extern int rcu_expedited; /* from sysctl */
67 module_param(rcu_expedited, int, 0);
68 extern int rcu_normal; /* from sysctl */
69 module_param(rcu_normal, int, 0);
70 static int rcu_normal_after_boot;
71 module_param(rcu_normal_after_boot, int, 0);
72 #endif /* #ifndef CONFIG_TINY_RCU */
74 #ifdef CONFIG_DEBUG_LOCK_ALLOC
76 * rcu_read_lock_sched_held() - might we be in RCU-sched read-side critical section?
78 * If CONFIG_DEBUG_LOCK_ALLOC is selected, returns nonzero iff in an
79 * RCU-sched read-side critical section. In absence of
80 * CONFIG_DEBUG_LOCK_ALLOC, this assumes we are in an RCU-sched read-side
81 * critical section unless it can prove otherwise. Note that disabling
82 * of preemption (including disabling irqs) counts as an RCU-sched
83 * read-side critical section. This is useful for debug checks in functions
84 * that required that they be called within an RCU-sched read-side
87 * Check debug_lockdep_rcu_enabled() to prevent false positives during boot
88 * and while lockdep is disabled.
90 * Note that if the CPU is in the idle loop from an RCU point of
91 * view (ie: that we are in the section between rcu_idle_enter() and
92 * rcu_idle_exit()) then rcu_read_lock_held() returns false even if the CPU
93 * did an rcu_read_lock(). The reason for this is that RCU ignores CPUs
94 * that are in such a section, considering these as in extended quiescent
95 * state, so such a CPU is effectively never in an RCU read-side critical
96 * section regardless of what RCU primitives it invokes. This state of
97 * affairs is required --- we need to keep an RCU-free window in idle
98 * where the CPU may possibly enter into low power mode. This way we can
99 * notice an extended quiescent state to other CPUs that started a grace
100 * period. Otherwise we would delay any grace period as long as we run in
103 * Similarly, we avoid claiming an SRCU read lock held if the current
106 int rcu_read_lock_sched_held(void)
108 int lockdep_opinion = 0;
110 if (!debug_lockdep_rcu_enabled())
112 if (!rcu_is_watching())
114 if (!rcu_lockdep_current_cpu_online())
117 lockdep_opinion = lock_is_held(&rcu_sched_lock_map);
118 return lockdep_opinion || !preemptible();
120 EXPORT_SYMBOL(rcu_read_lock_sched_held);
123 #ifndef CONFIG_TINY_RCU
126 * Should expedited grace-period primitives always fall back to their
127 * non-expedited counterparts? Intended for use within RCU. Note
128 * that if the user specifies both rcu_expedited and rcu_normal, then
129 * rcu_normal wins. (Except during the time period during boot from
130 * when the first task is spawned until the rcu_set_runtime_mode()
131 * core_initcall() is invoked, at which point everything is expedited.)
133 bool rcu_gp_is_normal(void)
135 return READ_ONCE(rcu_normal) &&
136 rcu_scheduler_active != RCU_SCHEDULER_INIT;
138 EXPORT_SYMBOL_GPL(rcu_gp_is_normal);
140 static atomic_t rcu_expedited_nesting = ATOMIC_INIT(1);
143 * Should normal grace-period primitives be expedited? Intended for
144 * use within RCU. Note that this function takes the rcu_expedited
145 * sysfs/boot variable and rcu_scheduler_active into account as well
146 * as the rcu_expedite_gp() nesting. So looping on rcu_unexpedite_gp()
147 * until rcu_gp_is_expedited() returns false is a -really- bad idea.
149 bool rcu_gp_is_expedited(void)
151 return rcu_expedited || atomic_read(&rcu_expedited_nesting) ||
152 rcu_scheduler_active == RCU_SCHEDULER_INIT;
154 EXPORT_SYMBOL_GPL(rcu_gp_is_expedited);
157 * rcu_expedite_gp - Expedite future RCU grace periods
159 * After a call to this function, future calls to synchronize_rcu() and
160 * friends act as the corresponding synchronize_rcu_expedited() function
161 * had instead been called.
163 void rcu_expedite_gp(void)
165 atomic_inc(&rcu_expedited_nesting);
167 EXPORT_SYMBOL_GPL(rcu_expedite_gp);
170 * rcu_unexpedite_gp - Cancel prior rcu_expedite_gp() invocation
172 * Undo a prior call to rcu_expedite_gp(). If all prior calls to
173 * rcu_expedite_gp() are undone by a subsequent call to rcu_unexpedite_gp(),
174 * and if the rcu_expedited sysfs/boot parameter is not set, then all
175 * subsequent calls to synchronize_rcu() and friends will return to
176 * their normal non-expedited behavior.
178 void rcu_unexpedite_gp(void)
180 atomic_dec(&rcu_expedited_nesting);
182 EXPORT_SYMBOL_GPL(rcu_unexpedite_gp);
185 * Inform RCU of the end of the in-kernel boot sequence.
187 void rcu_end_inkernel_boot(void)
190 if (rcu_normal_after_boot)
191 WRITE_ONCE(rcu_normal, 1);
194 #endif /* #ifndef CONFIG_TINY_RCU */
197 * Test each non-SRCU synchronous grace-period wait API. This is
198 * useful just after a change in mode for these primitives, and
201 void rcu_test_sync_prims(void)
203 if (!IS_ENABLED(CONFIG_PROVE_RCU))
206 synchronize_rcu_expedited();
209 #if !defined(CONFIG_TINY_RCU) || defined(CONFIG_SRCU)
212 * Switch to run-time mode once RCU has fully initialized.
214 static int __init rcu_set_runtime_mode(void)
216 rcu_test_sync_prims();
217 rcu_scheduler_active = RCU_SCHEDULER_RUNNING;
218 rcu_test_sync_prims();
221 core_initcall(rcu_set_runtime_mode);
223 #endif /* #if !defined(CONFIG_TINY_RCU) || defined(CONFIG_SRCU) */
225 #ifdef CONFIG_DEBUG_LOCK_ALLOC
226 static struct lock_class_key rcu_lock_key;
227 struct lockdep_map rcu_lock_map =
228 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock", &rcu_lock_key);
229 EXPORT_SYMBOL_GPL(rcu_lock_map);
231 static struct lock_class_key rcu_bh_lock_key;
232 struct lockdep_map rcu_bh_lock_map =
233 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_bh", &rcu_bh_lock_key);
234 EXPORT_SYMBOL_GPL(rcu_bh_lock_map);
236 static struct lock_class_key rcu_sched_lock_key;
237 struct lockdep_map rcu_sched_lock_map =
238 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_sched", &rcu_sched_lock_key);
239 EXPORT_SYMBOL_GPL(rcu_sched_lock_map);
241 static struct lock_class_key rcu_callback_key;
242 struct lockdep_map rcu_callback_map =
243 STATIC_LOCKDEP_MAP_INIT("rcu_callback", &rcu_callback_key);
244 EXPORT_SYMBOL_GPL(rcu_callback_map);
246 int notrace debug_lockdep_rcu_enabled(void)
248 return rcu_scheduler_active != RCU_SCHEDULER_INACTIVE && debug_locks &&
249 current->lockdep_recursion == 0;
251 EXPORT_SYMBOL_GPL(debug_lockdep_rcu_enabled);
254 * rcu_read_lock_held() - might we be in RCU read-side critical section?
256 * If CONFIG_DEBUG_LOCK_ALLOC is selected, returns nonzero iff in an RCU
257 * read-side critical section. In absence of CONFIG_DEBUG_LOCK_ALLOC,
258 * this assumes we are in an RCU read-side critical section unless it can
259 * prove otherwise. This is useful for debug checks in functions that
260 * require that they be called within an RCU read-side critical section.
262 * Checks debug_lockdep_rcu_enabled() to prevent false positives during boot
263 * and while lockdep is disabled.
265 * Note that rcu_read_lock() and the matching rcu_read_unlock() must
266 * occur in the same context, for example, it is illegal to invoke
267 * rcu_read_unlock() in process context if the matching rcu_read_lock()
268 * was invoked from within an irq handler.
270 * Note that rcu_read_lock() is disallowed if the CPU is either idle or
271 * offline from an RCU perspective, so check for those as well.
273 int rcu_read_lock_held(void)
275 if (!debug_lockdep_rcu_enabled())
277 if (!rcu_is_watching())
279 if (!rcu_lockdep_current_cpu_online())
281 return lock_is_held(&rcu_lock_map);
283 EXPORT_SYMBOL_GPL(rcu_read_lock_held);
286 * rcu_read_lock_bh_held() - might we be in RCU-bh read-side critical section?
288 * Check for bottom half being disabled, which covers both the
289 * CONFIG_PROVE_RCU and not cases. Note that if someone uses
290 * rcu_read_lock_bh(), but then later enables BH, lockdep (if enabled)
291 * will show the situation. This is useful for debug checks in functions
292 * that require that they be called within an RCU read-side critical
295 * Check debug_lockdep_rcu_enabled() to prevent false positives during boot.
297 * Note that rcu_read_lock_bh() is disallowed if the CPU is either idle or
298 * offline from an RCU perspective, so check for those as well.
300 int rcu_read_lock_bh_held(void)
302 if (!debug_lockdep_rcu_enabled())
304 if (!rcu_is_watching())
306 if (!rcu_lockdep_current_cpu_online())
308 return in_softirq() || irqs_disabled();
310 EXPORT_SYMBOL_GPL(rcu_read_lock_bh_held);
312 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
315 * wakeme_after_rcu() - Callback function to awaken a task after grace period
316 * @head: Pointer to rcu_head member within rcu_synchronize structure
318 * Awaken the corresponding task now that a grace period has elapsed.
320 void wakeme_after_rcu(struct rcu_head *head)
322 struct rcu_synchronize *rcu;
324 rcu = container_of(head, struct rcu_synchronize, head);
325 complete(&rcu->completion);
327 EXPORT_SYMBOL_GPL(wakeme_after_rcu);
329 void __wait_rcu_gp(bool checktiny, int n, call_rcu_func_t *crcu_array,
330 struct rcu_synchronize *rs_array)
335 /* Initialize and register callbacks for each crcu_array element. */
336 for (i = 0; i < n; i++) {
338 (crcu_array[i] == call_rcu)) {
342 init_rcu_head_on_stack(&rs_array[i].head);
343 init_completion(&rs_array[i].completion);
344 for (j = 0; j < i; j++)
345 if (crcu_array[j] == crcu_array[i])
348 (crcu_array[i])(&rs_array[i].head, wakeme_after_rcu);
351 /* Wait for all callbacks to be invoked. */
352 for (i = 0; i < n; i++) {
354 (crcu_array[i] == call_rcu))
356 for (j = 0; j < i; j++)
357 if (crcu_array[j] == crcu_array[i])
360 wait_for_completion(&rs_array[i].completion);
361 destroy_rcu_head_on_stack(&rs_array[i].head);
364 EXPORT_SYMBOL_GPL(__wait_rcu_gp);
366 #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
367 void init_rcu_head(struct rcu_head *head)
369 debug_object_init(head, &rcuhead_debug_descr);
371 EXPORT_SYMBOL_GPL(init_rcu_head);
373 void destroy_rcu_head(struct rcu_head *head)
375 debug_object_free(head, &rcuhead_debug_descr);
377 EXPORT_SYMBOL_GPL(destroy_rcu_head);
379 static bool rcuhead_is_static_object(void *addr)
385 * init_rcu_head_on_stack() - initialize on-stack rcu_head for debugobjects
386 * @head: pointer to rcu_head structure to be initialized
388 * This function informs debugobjects of a new rcu_head structure that
389 * has been allocated as an auto variable on the stack. This function
390 * is not required for rcu_head structures that are statically defined or
391 * that are dynamically allocated on the heap. This function has no
392 * effect for !CONFIG_DEBUG_OBJECTS_RCU_HEAD kernel builds.
394 void init_rcu_head_on_stack(struct rcu_head *head)
396 debug_object_init_on_stack(head, &rcuhead_debug_descr);
398 EXPORT_SYMBOL_GPL(init_rcu_head_on_stack);
401 * destroy_rcu_head_on_stack() - destroy on-stack rcu_head for debugobjects
402 * @head: pointer to rcu_head structure to be initialized
404 * This function informs debugobjects that an on-stack rcu_head structure
405 * is about to go out of scope. As with init_rcu_head_on_stack(), this
406 * function is not required for rcu_head structures that are statically
407 * defined or that are dynamically allocated on the heap. Also as with
408 * init_rcu_head_on_stack(), this function has no effect for
409 * !CONFIG_DEBUG_OBJECTS_RCU_HEAD kernel builds.
411 void destroy_rcu_head_on_stack(struct rcu_head *head)
413 debug_object_free(head, &rcuhead_debug_descr);
415 EXPORT_SYMBOL_GPL(destroy_rcu_head_on_stack);
417 struct debug_obj_descr rcuhead_debug_descr = {
419 .is_static_object = rcuhead_is_static_object,
421 EXPORT_SYMBOL_GPL(rcuhead_debug_descr);
422 #endif /* #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
424 #if defined(CONFIG_TREE_RCU) || defined(CONFIG_PREEMPT_RCU) || defined(CONFIG_RCU_TRACE)
425 void do_trace_rcu_torture_read(const char *rcutorturename, struct rcu_head *rhp,
427 unsigned long c_old, unsigned long c)
429 trace_rcu_torture_read(rcutorturename, rhp, secs, c_old, c);
431 EXPORT_SYMBOL_GPL(do_trace_rcu_torture_read);
433 #define do_trace_rcu_torture_read(rcutorturename, rhp, secs, c_old, c) \
437 #ifdef CONFIG_RCU_STALL_COMMON
439 #ifdef CONFIG_PROVE_RCU
440 #define RCU_STALL_DELAY_DELTA (5 * HZ)
442 #define RCU_STALL_DELAY_DELTA 0
445 int rcu_cpu_stall_suppress __read_mostly; /* 1 = suppress stall warnings. */
446 EXPORT_SYMBOL_GPL(rcu_cpu_stall_suppress);
447 static int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT;
449 module_param(rcu_cpu_stall_suppress, int, 0644);
450 module_param(rcu_cpu_stall_timeout, int, 0644);
452 int rcu_jiffies_till_stall_check(void)
454 int till_stall_check = READ_ONCE(rcu_cpu_stall_timeout);
457 * Limit check must be consistent with the Kconfig limits
458 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
460 if (till_stall_check < 3) {
461 WRITE_ONCE(rcu_cpu_stall_timeout, 3);
462 till_stall_check = 3;
463 } else if (till_stall_check > 300) {
464 WRITE_ONCE(rcu_cpu_stall_timeout, 300);
465 till_stall_check = 300;
467 return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
469 EXPORT_SYMBOL_GPL(rcu_jiffies_till_stall_check);
471 void rcu_sysrq_start(void)
473 if (!rcu_cpu_stall_suppress)
474 rcu_cpu_stall_suppress = 2;
477 void rcu_sysrq_end(void)
479 if (rcu_cpu_stall_suppress == 2)
480 rcu_cpu_stall_suppress = 0;
483 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
485 rcu_cpu_stall_suppress = 1;
489 static struct notifier_block rcu_panic_block = {
490 .notifier_call = rcu_panic,
493 static int __init check_cpu_stall_init(void)
495 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
498 early_initcall(check_cpu_stall_init);
500 #endif /* #ifdef CONFIG_RCU_STALL_COMMON */
502 #ifdef CONFIG_TASKS_RCU
505 * Simple variant of RCU whose quiescent states are voluntary context
506 * switch, cond_resched_rcu_qs(), user-space execution, and idle.
507 * As such, grace periods can take one good long time. There are no
508 * read-side primitives similar to rcu_read_lock() and rcu_read_unlock()
509 * because this implementation is intended to get the system into a safe
510 * state for some of the manipulations involved in tracing and the like.
511 * Finally, this implementation does not support high call_rcu_tasks()
512 * rates from multiple CPUs. If this is required, per-CPU callback lists
516 /* Global list of callbacks and associated lock. */
517 static struct rcu_head *rcu_tasks_cbs_head;
518 static struct rcu_head **rcu_tasks_cbs_tail = &rcu_tasks_cbs_head;
519 static DECLARE_WAIT_QUEUE_HEAD(rcu_tasks_cbs_wq);
520 static DEFINE_RAW_SPINLOCK(rcu_tasks_cbs_lock);
522 /* Track exiting tasks in order to allow them to be waited for. */
523 DEFINE_STATIC_SRCU(tasks_rcu_exit_srcu);
525 /* Control stall timeouts. Disable with <= 0, otherwise jiffies till stall. */
526 #define RCU_TASK_STALL_TIMEOUT (HZ * 60 * 10)
527 static int rcu_task_stall_timeout __read_mostly = RCU_TASK_STALL_TIMEOUT;
528 module_param(rcu_task_stall_timeout, int, 0644);
530 static struct task_struct *rcu_tasks_kthread_ptr;
533 * call_rcu_tasks() - Queue an RCU for invocation task-based grace period
534 * @rhp: structure to be used for queueing the RCU updates.
535 * @func: actual callback function to be invoked after the grace period
537 * The callback function will be invoked some time after a full grace
538 * period elapses, in other words after all currently executing RCU
539 * read-side critical sections have completed. call_rcu_tasks() assumes
540 * that the read-side critical sections end at a voluntary context
541 * switch (not a preemption!), cond_resched_rcu_qs(), entry into idle,
542 * or transition to usermode execution. As such, there are no read-side
543 * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
544 * this primitive is intended to determine that all tasks have passed
545 * through a safe state, not so much for data-strcuture synchronization.
547 * See the description of call_rcu() for more detailed information on
548 * memory ordering guarantees.
550 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func)
557 raw_spin_lock_irqsave(&rcu_tasks_cbs_lock, flags);
558 needwake = !rcu_tasks_cbs_head;
559 *rcu_tasks_cbs_tail = rhp;
560 rcu_tasks_cbs_tail = &rhp->next;
561 raw_spin_unlock_irqrestore(&rcu_tasks_cbs_lock, flags);
562 /* We can't create the thread unless interrupts are enabled. */
563 if (needwake && READ_ONCE(rcu_tasks_kthread_ptr))
564 wake_up(&rcu_tasks_cbs_wq);
566 EXPORT_SYMBOL_GPL(call_rcu_tasks);
569 * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
571 * Control will return to the caller some time after a full rcu-tasks
572 * grace period has elapsed, in other words after all currently
573 * executing rcu-tasks read-side critical sections have elapsed. These
574 * read-side critical sections are delimited by calls to schedule(),
575 * cond_resched_tasks_rcu_qs(), idle execution, userspace execution, calls
576 * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
578 * This is a very specialized primitive, intended only for a few uses in
579 * tracing and other situations requiring manipulation of function
580 * preambles and profiling hooks. The synchronize_rcu_tasks() function
581 * is not (yet) intended for heavy use from multiple CPUs.
583 * Note that this guarantee implies further memory-ordering guarantees.
584 * On systems with more than one CPU, when synchronize_rcu_tasks() returns,
585 * each CPU is guaranteed to have executed a full memory barrier since the
586 * end of its last RCU-tasks read-side critical section whose beginning
587 * preceded the call to synchronize_rcu_tasks(). In addition, each CPU
588 * having an RCU-tasks read-side critical section that extends beyond
589 * the return from synchronize_rcu_tasks() is guaranteed to have executed
590 * a full memory barrier after the beginning of synchronize_rcu_tasks()
591 * and before the beginning of that RCU-tasks read-side critical section.
592 * Note that these guarantees include CPUs that are offline, idle, or
593 * executing in user mode, as well as CPUs that are executing in the kernel.
595 * Furthermore, if CPU A invoked synchronize_rcu_tasks(), which returned
596 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
597 * to have executed a full memory barrier during the execution of
598 * synchronize_rcu_tasks() -- even if CPU A and CPU B are the same CPU
599 * (but again only if the system has more than one CPU).
601 void synchronize_rcu_tasks(void)
603 /* Complain if the scheduler has not started. */
604 RCU_LOCKDEP_WARN(rcu_scheduler_active == RCU_SCHEDULER_INACTIVE,
605 "synchronize_rcu_tasks called too soon");
607 /* Wait for the grace period. */
608 wait_rcu_gp(call_rcu_tasks);
610 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks);
613 * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
615 * Although the current implementation is guaranteed to wait, it is not
616 * obligated to, for example, if there are no pending callbacks.
618 void rcu_barrier_tasks(void)
620 /* There is only one callback queue, so this is easy. ;-) */
621 synchronize_rcu_tasks();
623 EXPORT_SYMBOL_GPL(rcu_barrier_tasks);
625 /* See if tasks are still holding out, complain if so. */
626 static void check_holdout_task(struct task_struct *t,
627 bool needreport, bool *firstreport)
631 if (!READ_ONCE(t->rcu_tasks_holdout) ||
632 t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) ||
633 !READ_ONCE(t->on_rq) ||
634 (IS_ENABLED(CONFIG_NO_HZ_FULL) &&
635 !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) {
636 WRITE_ONCE(t->rcu_tasks_holdout, false);
637 list_del_init(&t->rcu_tasks_holdout_list);
641 rcu_request_urgent_qs_task(t);
645 pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
646 *firstreport = false;
649 pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
650 t, ".I"[is_idle_task(t)],
651 "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)],
652 t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout,
653 t->rcu_tasks_idle_cpu, cpu);
657 /* RCU-tasks kthread that detects grace periods and invokes callbacks. */
658 static int __noreturn rcu_tasks_kthread(void *arg)
661 struct task_struct *g, *t;
662 unsigned long lastreport;
663 struct rcu_head *list;
664 struct rcu_head *next;
665 LIST_HEAD(rcu_tasks_holdouts);
668 /* Run on housekeeping CPUs by default. Sysadm can move if desired. */
669 housekeeping_affine(current, HK_FLAG_RCU);
672 * Each pass through the following loop makes one check for
673 * newly arrived callbacks, and, if there are some, waits for
674 * one RCU-tasks grace period and then invokes the callbacks.
675 * This loop is terminated by the system going down. ;-)
679 /* Pick up any new callbacks. */
680 raw_spin_lock_irqsave(&rcu_tasks_cbs_lock, flags);
681 list = rcu_tasks_cbs_head;
682 rcu_tasks_cbs_head = NULL;
683 rcu_tasks_cbs_tail = &rcu_tasks_cbs_head;
684 raw_spin_unlock_irqrestore(&rcu_tasks_cbs_lock, flags);
686 /* If there were none, wait a bit and start over. */
688 wait_event_interruptible(rcu_tasks_cbs_wq,
690 if (!rcu_tasks_cbs_head) {
691 WARN_ON(signal_pending(current));
692 schedule_timeout_interruptible(HZ/10);
698 * Wait for all pre-existing t->on_rq and t->nvcsw
699 * transitions to complete. Invoking synchronize_rcu()
700 * suffices because all these transitions occur with
701 * interrupts disabled. Without this synchronize_rcu(),
702 * a read-side critical section that started before the
703 * grace period might be incorrectly seen as having started
704 * after the grace period.
706 * This synchronize_rcu() also dispenses with the
707 * need for a memory barrier on the first store to
708 * ->rcu_tasks_holdout, as it forces the store to happen
709 * after the beginning of the grace period.
714 * There were callbacks, so we need to wait for an
715 * RCU-tasks grace period. Start off by scanning
716 * the task list for tasks that are not already
717 * voluntarily blocked. Mark these tasks and make
718 * a list of them in rcu_tasks_holdouts.
721 for_each_process_thread(g, t) {
722 if (t != current && READ_ONCE(t->on_rq) &&
725 t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw);
726 WRITE_ONCE(t->rcu_tasks_holdout, true);
727 list_add(&t->rcu_tasks_holdout_list,
728 &rcu_tasks_holdouts);
734 * Wait for tasks that are in the process of exiting.
735 * This does only part of the job, ensuring that all
736 * tasks that were previously exiting reach the point
737 * where they have disabled preemption, allowing the
738 * later synchronize_rcu() to finish the job.
740 synchronize_srcu(&tasks_rcu_exit_srcu);
743 * Each pass through the following loop scans the list
744 * of holdout tasks, removing any that are no longer
745 * holdouts. When the list is empty, we are done.
747 lastreport = jiffies;
749 /* Start off with HZ/10 wait and slowly back off to 1 HZ wait*/
756 struct task_struct *t1;
758 if (list_empty(&rcu_tasks_holdouts))
761 /* Slowly back off waiting for holdouts */
762 schedule_timeout_interruptible(HZ/fract);
767 rtst = READ_ONCE(rcu_task_stall_timeout);
768 needreport = rtst > 0 &&
769 time_after(jiffies, lastreport + rtst);
771 lastreport = jiffies;
773 WARN_ON(signal_pending(current));
774 list_for_each_entry_safe(t, t1, &rcu_tasks_holdouts,
775 rcu_tasks_holdout_list) {
776 check_holdout_task(t, needreport, &firstreport);
782 * Because ->on_rq and ->nvcsw are not guaranteed
783 * to have a full memory barriers prior to them in the
784 * schedule() path, memory reordering on other CPUs could
785 * cause their RCU-tasks read-side critical sections to
786 * extend past the end of the grace period. However,
787 * because these ->nvcsw updates are carried out with
788 * interrupts disabled, we can use synchronize_rcu()
789 * to force the needed ordering on all such CPUs.
791 * This synchronize_rcu() also confines all
792 * ->rcu_tasks_holdout accesses to be within the grace
793 * period, avoiding the need for memory barriers for
794 * ->rcu_tasks_holdout accesses.
796 * In addition, this synchronize_rcu() waits for exiting
797 * tasks to complete their final preempt_disable() region
798 * of execution, cleaning up after the synchronize_srcu()
803 /* Invoke the callbacks. */
812 /* Paranoid sleep to keep this from entering a tight loop */
813 schedule_timeout_uninterruptible(HZ/10);
817 /* Spawn rcu_tasks_kthread() at core_initcall() time. */
818 static int __init rcu_spawn_tasks_kthread(void)
820 struct task_struct *t;
822 t = kthread_run(rcu_tasks_kthread, NULL, "rcu_tasks_kthread");
823 if (WARN_ONCE(IS_ERR(t), "%s: Could not start Tasks-RCU grace-period kthread, OOM is now expected behavior\n", __func__))
825 smp_mb(); /* Ensure others see full kthread. */
826 WRITE_ONCE(rcu_tasks_kthread_ptr, t);
829 core_initcall(rcu_spawn_tasks_kthread);
831 /* Do the srcu_read_lock() for the above synchronize_srcu(). */
832 void exit_tasks_rcu_start(void)
835 current->rcu_tasks_idx = __srcu_read_lock(&tasks_rcu_exit_srcu);
839 /* Do the srcu_read_unlock() for the above synchronize_srcu(). */
840 void exit_tasks_rcu_finish(void)
843 __srcu_read_unlock(&tasks_rcu_exit_srcu, current->rcu_tasks_idx);
847 #endif /* #ifdef CONFIG_TASKS_RCU */
849 #ifndef CONFIG_TINY_RCU
852 * Print any non-default Tasks RCU settings.
854 static void __init rcu_tasks_bootup_oddness(void)
856 #ifdef CONFIG_TASKS_RCU
857 if (rcu_task_stall_timeout != RCU_TASK_STALL_TIMEOUT)
858 pr_info("\tTasks-RCU CPU stall warnings timeout set to %d (rcu_task_stall_timeout).\n", rcu_task_stall_timeout);
860 pr_info("\tTasks RCU enabled.\n");
861 #endif /* #ifdef CONFIG_TASKS_RCU */
864 #endif /* #ifndef CONFIG_TINY_RCU */
866 #ifdef CONFIG_PROVE_RCU
869 * Early boot self test parameters.
871 static bool rcu_self_test;
872 module_param(rcu_self_test, bool, 0444);
874 static int rcu_self_test_counter;
876 static void test_callback(struct rcu_head *r)
878 rcu_self_test_counter++;
879 pr_info("RCU test callback executed %d\n", rcu_self_test_counter);
882 DEFINE_STATIC_SRCU(early_srcu);
884 static void early_boot_test_call_rcu(void)
886 static struct rcu_head head;
887 static struct rcu_head shead;
889 call_rcu(&head, test_callback);
890 if (IS_ENABLED(CONFIG_SRCU))
891 call_srcu(&early_srcu, &shead, test_callback);
894 void rcu_early_boot_tests(void)
896 pr_info("Running RCU self tests\n");
899 early_boot_test_call_rcu();
900 rcu_test_sync_prims();
903 static int rcu_verify_early_boot_tests(void)
906 int early_boot_test_counter = 0;
909 early_boot_test_counter++;
911 if (IS_ENABLED(CONFIG_SRCU)) {
912 early_boot_test_counter++;
913 srcu_barrier(&early_srcu);
916 if (rcu_self_test_counter != early_boot_test_counter) {
923 late_initcall(rcu_verify_early_boot_tests);
925 void rcu_early_boot_tests(void) {}
926 #endif /* CONFIG_PROVE_RCU */
928 #ifndef CONFIG_TINY_RCU
931 * Print any significant non-default boot-time settings.
933 void __init rcupdate_announce_bootup_oddness(void)
936 pr_info("\tNo expedited grace period (rcu_normal).\n");
937 else if (rcu_normal_after_boot)
938 pr_info("\tNo expedited grace period (rcu_normal_after_boot).\n");
939 else if (rcu_expedited)
940 pr_info("\tAll grace periods are expedited (rcu_expedited).\n");
941 if (rcu_cpu_stall_suppress)
942 pr_info("\tRCU CPU stall warnings suppressed (rcu_cpu_stall_suppress).\n");
943 if (rcu_cpu_stall_timeout != CONFIG_RCU_CPU_STALL_TIMEOUT)
944 pr_info("\tRCU CPU stall warnings timeout set to %d (rcu_cpu_stall_timeout).\n", rcu_cpu_stall_timeout);
945 rcu_tasks_bootup_oddness();
948 #endif /* #ifndef CONFIG_TINY_RCU */